Abstract: Kawin Ethayarajh, David Duvenaud, Graeme Hirst. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.
2019.
Introduction
Word embeddings like Word2Vec (SGNS) and GloVe are fundamental to modern NLP, yet they often inherit and even amplify human prejudices found in training data—for instance, associating "doctor" more strongly with "man" than "woman." While debiasing methods exist, the field has lacked a rigorous theoretical foundation explaining why they work. This paper provides that framework, proving that subspace projection is mathematically sound and revealing that current bias metrics, such as the Word Embedding Association Test (WEAT), are statistically flawed.
Methods
The authors introduce three major technical contributions to the study of embedding bias:
- The Debiasing Theorem: The paper proves that for models performing implicit matrix factorization (like SGNS and GloVe), subtracting a word vector's projection onto a "bias subspace" is mathematically equivalent to training the model on a perfectly unbiased corpus.
- Relational Inner Product Association (RIPA): A new metric for measuring bias that uses the inner product of a word vector and a "relation vector" (the first principal component of pairs like man-woman). RIPA is highly interpretable, relating directly to the log-odds of a word co-occurring with one gender versus another.
- Unsupervised Word Classification: To prevent "over-debiasing" that destroys legitimate semantic relationships (e.g., king vs. queen), the authors developed an unsupervised method to distinguish "gender-appropriate" words from "gender-biased" ones (e.g., nurse).
Key Insights: The Amplification Effect
A critical finding of this research is that embedding models do not merely mirror the training data; they often amplify existing stereotypes. This occurs because words like "nurse" and "female" share similar contexts beyond simple co-occurrence. In low-dimensional vector spaces, this "second-order similarity" pushes these vectors closer together than their raw corpus statistics would suggest.
Furthermore, the authors identify that the widely used WEAT metric is overly sensitive to word frequency. If "male" attributes appear more frequently than "female" ones in a corpus, WEAT may incorrectly flag neutral words as biased. They also demonstrate that normalizing word vectors to unit length before debiasing—a common practice—is counterproductive, as vector length contains vital frequency information.
Key Results
The proposed methods offer significant improvements in both interpretability and performance:
- Quantifying Amplification: For gender-biased words, "genderedness" increased by an average of 0.197 from the corpus to the embedding space. Gender-appropriate words saw an even larger increase of 0.291.
- Preserving Legitimate Analogies: The unsupervised debiasing method proved far superior to previous supervised approaches. It preserved 94.9% of gender-appropriate analogies (like king:queen), whereas the previous state-of-the-art preserved only 16.5%.
Limitations and Future Work
The core proof assumes "perfect reconstruction"—that the embedding model perfectly factorizes the co-occurrence matrix. In practice, reconstruction noise may slightly affect the theoretical equivalence guarantee. Additionally, while this study focuses on gender, the authors anticipate these findings will extend to other social biases (race, religion, etc.). Future research could also investigate whether undesirable associations exist within non-linear manifolds of the embedding space.
Summary updated